Identifying and Understanding Differential Transcriptor Binding

Transcription

1 Identifying and Understanding Differential Transcriptor Binding : Computational Genomics David Koes Yong Lu

2 Motivation Under different conditions, a transcription factor binds to different genes Why? YPD (rich media) GLN3 When? rapamycin GLN3 Where? GLN1 DAL David Koes and Yong Lu 1

3 Example: Difference Graph Nodes higher ypd expr higher rap expr Edges rap binding only ypd binding only ypd & rap binding protein interaction 2004 David Koes and Yong Lu 2

4 Goals Can we fill in the missing values of a binding experiment? Can we predict all the values of a binding experiment? Can we explain the differences in binding? 2004 David Koes and Yong Lu 3

5 Approach: Classification Given a transcription factor/gene pair will there be binding under rapamycin? Data GLN3? DAL1 Classifier GLN3 DAL David Koes and Yong Lu 4

6 Outline Data Sources Feature Selection Classifiers Results 2004 David Koes and Yong Lu 5

7 Data: Saccharomyces cerevisiae Expression ypd and rapamycin micro-array data Binding genome wide location analysis YPD: Rapamycin: Protein Interaction two-hybrid method David Koes and Yong Lu 6

8 Outline Data Sources Feature Selection sparse and precise dense and aggregate Classifiers Results 2004 David Koes and Yong Lu 7

9 Features? GLN3? DAL1 Classifier GLN3 DAL David Koes and Yong Lu 8

10 Features Expression, binding, protein data not features global values not dependant upon a given edge Must exploit topology of data networks Expression Binding GLN3? Proteins DAL David Koes and Yong Lu 9

11 Outline Data Sources Feature Selection sparse and precise dense and aggregate Classifiers Results 2004 David Koes and Yong Lu 10

12 Sparse and Precise Several attributes for every gene binding pvalue for gene with factor/target expression of gene if gene can bind factor/target expression of gene if factor/target can bind gene expression of gene if protein interaction with factor/target exists expression of gene if gene is factor/target 2004 David Koes and Yong Lu 11

13 Example Nonzero attributes (16) factor_binds_target_rap.6 factor_binds_target_ypd.001 factor_binds_c_ypd.001 B_binds_target_ypd.002 factor_self_a_ypd 2.0 factor_1down_c_ypd -2.0 factor_1pp_b_ypd -1.0 factor_self_a_rap 1.0 factor_1down_c_rap 2.0 factor_1pp_b_rap 1.0 target_self_c_ypd -2.0 target_1up_a_ypd 2.0 target_1up_b_ypd -1.0 target_self_c_rap 2.0 target_1up_a_rap 1.0 target_1up_b_rap 1.0 Zero attributes (46) factor_binds_a_ypd factor_binds_b_ypd target_binds_a_ypd target_binds_b_ypd <etc.> ypd: 2.0 rap: 1.0 A pval:.001 pval:.6 C B ypd: -2.0 rap: 2.0 ypd: -1.0 rap: 1.0 pval: David Koes and Yong Lu 12

14 Pros and Cons Pros Precisely captures all the data Sparse dataset results in compact representation Solvers can take advantage of sparseness Cons Susceptible to over-fitting Huge number of attributes Solvers require binary attributes 2004 David Koes and Yong Lu 13

15 Outline Data Sources Feature Selection sparse and precise dense and aggregate Classifiers Results 2004 David Koes and Yong Lu 14

16 Dense and Aggregate Use averages of data based on topological relationship in network genes that can bind factor/target genes that factor/target can bind genes with protein interactions with factor/target YPD binding data 2004 David Koes and Yong Lu 15

17 Example Nonzero attributes (12) rap_bind.6 ypd_bind factor_expr_ypd 2.0 factor_expr_rap 1.0 target_expr_ypd -2.0 target_expr_rap 2.0 target_ave_expr_up_ypd 0.5 target_ave_expr_up_rap 1.0 factor_ave_expr_down_ypd -2.0 factor_ave_expr_down_rap 2.0 factor_ave_expr_pp_ypd -1.0 factor_ave_expr_pp_rap 1.0 Zero attributes (6) target_ave_expr_down_ypd 0 target_ave_expr_down_rap 0 target_ave_expr_pp_ypd 0 target_ave_expr_pp_rap 0 factor_ave_expr_up_ypd 0 factor_ave_expr_up_rap 0 ypd: 2.0 rap: 1.0 A pval:.001 pval:.6 B ypd: -1.0 rap: 1.0 pval:.002 C ypd: -2.0 rap: David Koes and Yong Lu 16

18 Pros and Cons Pros Small, constant, number of attributes Low penalty for adding additional attributes Cons Information lost 2004 David Koes and Yong Lu 17

19 Outline Data Sources Feature Selection Classifiers Results Logistic Regression K Nearest Neighbor Naïve Bayes Learned Bayes Net 2004 David Koes and Yong Lu 18

20 Logistic Regression Find β such that µ best approximate the training data outputs y where Solved with iterative re-weighted least squares Newton-Raphson µ = e ( β x i ) i ( ) 1+ e β x i 2004 David Koes and Yong Lu 19

21 K Nearest Neighbors Classify a point based on value of training points close by in attribute space 2004 David Koes and Yong Lu 20

22 Naïve Bayes Makes simplifying assumption that attributes are conditional independent given class Uses training data to estimate conditional probabilities Classifies based on what class assignment maximizes joint probability 2004 David Koes and Yong Lu 21

23 Learned Bayes Net Use training data to find a good network of conditional dependencies 2004 David Koes and Yong Lu 22

24 Outline Data Sources Feature Selection Classifiers Results 2004 David Koes and Yong Lu 23

25 Tools Auton Fast Classifiers Bayes Net Inference BNT/Matlab David Koes and Yong Lu 24

26 Goals Can we fill in the missing values of a binding experiment? Can we predict all the values of a binding experiment? Can we explain the differences in binding? 2004 David Koes and Yong Lu 25

27 Evaluation Use data from all 12 transcription factors Training set all edges with binding in either condition randomly selected nonbinding edges k-fold validation use 1/k th of data as test set simulates missing values 2004 David Koes and Yong Lu 26

28 ROC Curve: Sparse Naïve Bayes 2004 David Koes and Yong Lu 27

29 S p a r s e N. Bayes KNN LR D e n s e 2004 David Koes and Yong Lu 28

30 K-Folds AUC Area Under Curve Naïve Bayes Sparse Naïve Bayes Dense KNN Sparse KNN Dense LR Sparse LR Dense Number Folds 2004 David Koes and Yong Lu 29

31 8-Fold, Single Factor Area Under Curve DAL81 DAL82 FHL1 GAT1 GCN4 GLN3 Transcription Factor HAP2 MSN2 MSN4 RTG1 RTG3 UGA3 Naïve Bayes Sparse Naïve Bayes Dense KNN Sparse KNN Dense LR Sparse LR Dense 2004 David Koes and Yong Lu 30

32 Goals Can we fill in the missing values of a binding experiment? Can we predict all the values of a binding experiment? Can we explain the differences in binding? 2004 David Koes and Yong Lu 31

33 Evaluation Training set full data for 11 transcription factors Test set full data of remaining transcription factor 2004 David Koes and Yong Lu 32

34 ROC Curves: Sparse N. Bayes BAD! 2004 David Koes and Yong Lu 33

35 ROC Curves: Dense LR 2004 David Koes and Yong Lu 34

36 S p a r s e N. Bayes KNN LR D e n s e 2004 David Koes and Yong Lu 35

37 AUC: Leave One Out Area Under Curve DAL81 DAL82 FHL1 GAT1 GCN4 GLN3 Transcription Factor HAP2 MSN2 MSN4 RTG1 RTG3 UGA3 Naïve Bayes Sparse Naïve Bayes Dense KNN Sparse KNN Dense LR Sparse LR Dense 2004 David Koes and Yong Lu 36

38 Unknown Transcription Factors Rapamycin data for only 12 factors YPD data for 106 factors What is predicted for additional factors? Use sparse LR Only consider already binding YPD edges 2004 David Koes and Yong Lu 37

39 Top 20 Most Differing Factors 4 1 FHL1 GAT1 DAL82 UGA3 RAP1 MSN4 MSN2 ABF1 HAP2 CIN5 94% 93% 91% 90% 88% 82% 82% 79% 67% 64% 1 DAL81 GLN3 RTG1 REB1 MCM1 FKH1 RCS1 SWI4 RTG3 FZF1 PubMed Hits 62% 61% 60% 59% 48% 47% 46% 44% 43% 41% 2004 David Koes and Yong Lu 38

40 Goals Can we fill in the missing values of a binding experiment? Can we predict all the values of a binding experiment? Can we explain the differences in binding? 2004 David Koes and Yong Lu 39

41 Learned Bayes Network Simple classifiers may be successful but don t generate intuitive models Bayesian network might infer causality Find network that explains (dense) data well 2004 David Koes and Yong Lu 40

42 Learned Baysian Network 2004 David Koes and Yong Lu 41

43 Conclusion Classifiers very good at filling in missing values Classifiers can sometimes predict results of an experiment but sometimes way off Results may be used as guide to experimentation There may be some biological meaning within the classifier s model 2004 David Koes and Yong Lu 42